iShadow: The Computational Eyeglass

iShadow: The Computational Eyeglass
Addison Mayberry, Pan Hu, Christopher Salthouse, Benjamin Marlin, Deepak Ganesan
Introduction
iShadow Platform Prototype
The iShadow project aims to develop an end-to-end
system that includes novel ultra-low power
computational eyeglasses capable of detecting eye
movement and features of the external environment. It
will provide real-time sensing, processing, and inference
capability. Using this system, we will be able to provide
the fundamental knowledge base necessary to discover
patterns of human behavior and to leverage such
patterns to improve transportation and healthcare.
SD Storage
Microprocessor
(Hidden)
Eye-Facing Camera
Environment Camera
Sparse Eye Image Acquisition for Gaze Tracking
1. We collect training data of
a subject’s eye following a
dot on a monitor
• We posed gaze tracking as a
classification problem
• Our cameras are accessed pixel-by-pixel,
so if we are able to use fewer pixels we
conserve power and time
• By treating the pixels as our features,
we can do an analysis to determine
which pixels are providing the most
valuable information
3. We are able to get low gaze
prediction error with a small
number of pixels
2. To develop a sparse model, we
penalize the weights of pixels in a
neural network model
Configurable System Performance Tradeoffs
• The number of pixels selected is
controlled by the regularization
parameter, λ
• More pixels also implies a higher
energy cost, we can similarly
regulate energy consumption
• More pixels means higher
accuracy but longer latency
for prediction
• By adjusting λ we can select an
appropriate operating point on
this curve
Acknowledgements
For further info on the Sensors Lab:
Partial funding for the project came from the National Science
Foundation, NSF Grant: 1239341
For information on current work in the Sensors Lab, visit
http://sensors.cs.umass.edu